Grid Based Spherical CNN for Object Detection from Panoramic Images.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Spherical convolutional neural networks (SCNNs) show promise for spherical image classification.
• Conventional SCNNs face challenges in object detection due to location information loss and high bandwidth requirements.
• Existing methods like Faster R-CNN and SSD are not optimized for spherical data.

Purpose of the Study:

• To develop a novel grid-based spherical CNN (G-SCNN) for accurate object detection in spherical images.
• To address the limitations of existing SCNNs in preserving object location and managing bandwidth.
• To introduce a new dataset for evaluating spherical object detection models.

Main Methods:

• Transformation of spherical images to a conformal grid map for S2/SO(3) convolution.
• Utilization of a planar region proposal network (RPN) with rotation invariance data augmentation.
• Creation and annotation of the WHU panoramic dataset with 5636 objects from 600 images.

Main Results:

• The proposed G-SCNN effectively preserves object location information.
• The grid-based approach significantly reduces bandwidth requirements for spherical object detection.
• G-SCNN demonstrated superior performance compared to original SCNN, Faster R-CNN, and SSD on the WHU dataset.

Conclusions:

• G-SCNN offers a robust solution for object detection in spherical imagery.
• The method overcomes key limitations of previous spherical convolutional approaches.
• The WHU dataset provides a valuable resource for advancing spherical object detection research.